Does a smarter AI forecast help or hurt, depending on whether your teams work independently or are tightly linked together?

This explores whether an AI prediction is worth more or less depending on how separate or interlocked an organization's decisions are. Only one note in the collection answers this directly, but it gives a clear answer. Does AI adoption require rethinking how organizations coordinate? presents a formal model in which AI makes each individual task decision better and also makes those decisions vary more. If a decision stands alone, that variation does no harm, because you just get a better local call. If the decision is tied to others, so that a change in pricing affects inventory, which affects staffing, the variation becomes a coordination problem. A better prediction in one place can throw the rest of the system off unless the organization changes how its parts communicate. So modularity works in two directions. In a loosely connected organization you can drop an AI prediction into one spot and capture its value quickly. In a tightly connected one, most of the value (the 'synergistic returns' across connected decisions) only shows up after you redesign how those decisions fit together.

The consulting and practitioner notes describe the same thing from the field. Does agentic AI succeed through process redesign or workflow layering? says many agent deployments fail because companies add AI on top of existing workflows. That is the strategy that only works when the workflow is already modular. Does easier tool-building actually solve enterprise adoption problems? adds a less obvious barrier: once a useful AI tool affects work in several departments, adopting it becomes a decision for the whole organization, with all the timelines and politics that brings. Making the tool easier to build does nothing to remove that.

How should humans and AI systems divide decision-making tasks? offers a way to design for this. Puranam treats a human-AI pair as a small organization, where splitting up tasks and deciding how feedback flows are design choices. The point is that 'how modular are we?' doesn't have to be a fixed fact you put up with. You can decide where to draw the boundaries so that AI predictions land in places where they can be used.

The surprising part is that AI researchers have found the same pattern inside AI systems. Does separating planning from execution improve reasoning accuracy? shows that splitting a model into a separate planner and a separate executor beats a single monolithic model, partly because the two jobs stop interfering with each other. Do single agents always hit organizational limits? argues that a single agent hits organizational limits that more capability can't fix. Can orchestration strategies boost diagnostic AI without better models? gives a concrete example: on medical diagnosis cases, wrapping the same model in a structured arrangement of roles kept accuracy about the same or slightly higher and cut cost by roughly 70%. In each case the value came from how the parts were organized, not from a better predictor. Economists and AI engineers have reached the same conclusion from opposite sides: how good a prediction is matters less than whether the structure around it can absorb that prediction.

One honest limit: the collection doesn't contain empirical studies that measure AI's value across organizations with more or less modular structures. What's here is one formal model, practitioner arguments, and parallels from AI system design. Those parallels point the same way, but they are not direct evidence.